For convolution-crash.svg a linux-specific baseline seems to be needed.
Created attachment 62929 [details] Patch
Can you please discribe in the changelog why the result chanhes? Just a note to the changeset that causes this changes and the comment that you rebaseline chromium linux. You don't need to mention that the result comes directly from the bot.
jorlow fixed the immediate failure here (I should have put the windows baseline not the linux one) and I'm going to leave this open so that I can figure out why there are multiple baselines at all.
I figured it out, I was sloppy when I wrote the test case and the bad convolution is actually on the screen at 0,0. I should have hidden it under the gray box so that we could have the same result on all platforms unaffected by how we choose to truncate for out of range values during the convolution. However as this is all passing I'm inclined to leave it alone.